OpenAI's Codex started as a tool to help software engineers write code faster. Over the past few months, it has been quietly building toward something larger: an agent that can operate software on your behalf, the way a capable junior employee might.
The latest step is browser control. Codex can now connect to Chrome at a deep level, reading what is happening inside a live web page: network traffic, console logs, page structure, performance data. It can also rewrite parts of a page on the fly, changing colors, spacing, and fonts, or pulling structured data out of a site automatically. Think of it as giving an AI the same view a web developer gets when they right-click a page and hit "inspect," except the AI can also act on what it sees.
This is an opt-in feature, gated behind a toggle in Settings. Admins at organizations can turn it off entirely. It is not yet available in the European Economic Area, the UK, or Switzerland. And it is slow and unreliable at this stage: the browser access works, but often only after careful instructions and multiple attempts. This is genuinely early.
The reason it matters beyond developers is the trajectory. OpenAI published its own usage data this month: Codex has over 5 million weekly users, up roughly 400% since the start of the year. About 20% of those users are non-technical knowledge workers, and that segment is growing three times faster than the developer segment. Codex is already being used for things like reviewing messages and calendars, drafting slide decks, extracting data into spreadsheets, and mapping out business workflows. None of those require any coding knowledge.
The Ona acquisition fits directly into this picture. Ona, a German startup previously known as Gitpod, built secure cloud environments where software can run for hours or days inside a company's own infrastructure. The current limitation with any AI agent is that it stops when your session ends. Ona solves that. A task you give Codex on Monday morning could still be running, safely, on Tuesday. Once integrated, OpenAI says the combined product will target things like automated vulnerability scanning, codebase modernization, and long-running testing, but the underlying capability applies to any multi-step process that currently requires a person to babysit it in real time.
The competitive picture is sharp. Anthropic's Claude Code product has been growing fast, and Anthropic responded to OpenAI's Codex pricing moves by raising Claude Code's usage limits by 50%. Both companies are racing to be the agent platform that large enterprises trust with production-level work, not just experimentation.
For non-technical professionals, the practical signal is this: browser-operating AI is no longer theoretical, and the tools being built now are designed with you in mind, not just engineers. Repetitive web-based work, data gathering, monitoring, and reporting tasks are the near-term targets. The current version is slow and limited. But the pieces are assembling faster than most people outside the software industry have noticed.